С теорией ARC многоагентные LLM получают детерминистические гарантии безопасности

Qwen 2.5 7B Instruct agents running on a consumer 24GB GPU at a five-minute cadence produced auditable control trajectories for a dairy barn ventilation system, with every inter-agent conflict resolved deterministically—regardless of what the LLMs output themselves.

Mapping Feedback Loops to LLM Agents

The paper decomposes a process control problem using Advanced Regulatory Control (ARC) theory. Each feedback loop in the ARC chain maps to one specialized LLM operator agent. That agent carries the loop's core context: controlled variable, setpoint, chain priority, and selector kind. This bounds the task each model sees. A general-purpose LLM like Qwen 2.5 7B only needs to produce an operator action for its narrow slice of the system.

Two orchestrator variants sit above the agents. The first is a deterministic rule chain that implements MIN/MAX selectors and override paths as pure logic. The second uses a Claude-based LLM orchestrator at a slower tier. Both enforce the same safety property: every constraint conflict is resolved by the orchestrator, not by the individual agent outputs.

Deterministic Conflict Resolution via MIN/MAX Selectors

Control theory provides a discipline for decomposing a system into elements of contained scope, each defending one controlled variable. Conflicts between controlled variables are resolved by structural priority: MIN/MAX selector networks for CV-CV switching, and split-range (split-parallel) logic for MV-MV switching. The paper encapsulates this interaction logic inside the orchestrator, not the agents.

This design means the multi-agent system inherits the safety property of the ARC chain. Even if an LLM agent produces a wildly off-target action, the orchestrator's selector network overrides it. Auditing requires only checking the selector network's resolution, not tracing LLM reasoning.

Auditability and Practical Feasibility

Evaluated on a dairy-barn ventilation case over a 4-day mixed-season scenario, the approach produced auditable trajectories. Each control action is paired with an operator-voice rationale that supports a control campaign logbook. The entire system runs offline on a 24 GB consumer GPU—no cloud dependency, no special hardware.

Control theory's decades-old safety discipline is directly portable to LLM-based automation. Compliance audits become as straightforward as checking the selector network, not interpreting a black-box model's reasoning chain.

Source: A Systematic Approach to Multi-Agent AI from Advanced Regulatory Control Theory: Safe and Auditable LLM Operator Agents for Process Control
Domain: arxiv.org